EPN: Efficient and Balanced Network for Lightweight 2D Human Pose Estimation

With advances in deep learning, human pose estimation has seen significant progress. However, deep learning-based methods often increase parameter counts to boost accuracy, which leads to higher computational costs and restricts practical applications. In response, lightweight models for human pose estimation have been developed to lower computational demands, but they often suffer from reduced accuracy. To address this trade-off, we introduce the Efficient Pose Network (EPN), a novel 2D lightweight human pose estimation method. Our method incorporates innovative strategies in model architecture, training strategies, and keypoint detection. Extensive experiments demonstrate that EPN offers an outstanding balance of performance and efficiency, achieving high accuracy on the MPII dataset and SOTA results on the COCO dataset with identical parameter settings.